He was therefore quite bullish on NVIDIA's ability to use its design prowess to compete with Intel on both performance and power efficiency, using standard ARM cores and a foundry process. Huang's response was to point out that NVIDIA's PC chipsets were routinely top performers, so the company has plenty of experience with the kinds of system integration issues that I raised. all have a huge impact on real-world application performance, far and above that of the peak theoretical single-threaded performance of an individual core. My point was that these kinds of finicky issues like cache hierarchy design, choice of on-chip interconnect technology, floorplan layout, memory bus bandwidth, choice of multiprocessor interconnect topology, etc. As an example of the latter, I brought up AMD's problems with L2 cache latency, which crippled the performance of their server parts. Intel's track record in both of these areas is very good, in part because of the chipmaker's process leadership and in part because performance and power are increasingly sensitive to SoC- and system-level integration issues. ![]() ![]() Put differently, I suggested that if we move to a world where ISA really doesn't matter to the user-facing parts of the software stack (i.e., Windows and Flash both run on ARM, and Android and WP7 use just-in-time compilation for application binaries), then ARM products like NVIDIA's get to compete with Intel's chips on old-fashioned performance and power savings. Intel has been working on exactly this problem for over a decade. Specifically, I pointed out that attempts to do high-performance, power-efficient, out-of-order processors are not, in fact, even remotely new if we ignore the specific instruction set architecture (ISA). In the course of the evening, I tried to press Huang a bit on his second point. Both projects see NVIDIA pushing the ARM architecture to new heights of performance, and will let the chipmaker claim a set of "firsts" (i.e., first quad-core A9 part on the market, and first desktop-caliber ARM/GPU combo part). In his talk of breaking new ground, he seemed to be referring to both Kal-El (the recently unveiled quad-core Tegra 3 chip) and Project Denver as examples of NVIDIA making something brand new. The second reason that Huang gave for going with ARM was that he wanted the company to do something that had never been done before-to make a product that didn't yet exist. But he gave two reasons why the company opted not to go down that road. x86, he elaborated a bit on the rationale behind Project Denver. Huang was quite clear that NVIDIA could have chosen to produce an x86 processor-he described the licensing and technical problems associated with making an x86 CPU as "solvable" for NVIDIA. In the course of his discussion of ARM vs. Huang also heavily emphasized ARM as the future of not just NVIDIA, but the entire CPU business. Mapping the first two stages of NVIDIA's evolution to GPU's transition from a fixed-function ASIC to a full-blown math coprocessor makes sense, but the attempt to fit both Tegra and Tesla into the same "parallel computing" bucket felt a bit premature absent some news about Tegra finding its way into a cloud server product. NVIDIA 2.0 arrived with the advent of more fully programmable GPUs, and NVIDIA 3.0 will see the chipmaker transition into a "parallel computing company" that sells parallel number crunching hardware in every market from mobiles to servers. Huang began by dividing NVIDIA's history into three stages, the first of which-NVIDIA 1.0-was a period in which the company was a maker of fixed-function graphics products. Ultimately, we could even see Intel get back into the ARM market, a market where it had considerable success with its XScale line before betting the farm on x86. ![]() We'll recap Jen-Hsun's take on the processor and GPU markets, followed by a look at the implications of the trends he references for the future of Intel, the x86 instruction set architecture (ISA), ARM, and the CPU market as a whole. Jen-Hsun's remarks are worth looking at in some detail, as much for what they say about Intel as what they say about NVIDIA. At a dinner this week with members of the press, NVIDIA CEO Jen-Hsun Huang laid out his view of NVIDIA's past, present, and future in light of recent developments in the processor market.
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |